RPS 108 - Advanced imaging in head and neck tumours

Purpose:

Achieving good performance on real biomedical data that frequently contains imbalance characteristics remains a challenging task. This study aimed to investigate the impact of re-sampling techniques in imbalanced datasets for PET radiomics-based prognostication in head and neck (HNC) cancer patients.

PET-based radiomics analysis was performed in 166 patients (median age 49.3 years, 77.1% male) diagnosed with nasopharyngeal carcinoma (NPC) in our centre and 182 HNC patients from the TCIA database. Conventional PET and 15 robust texture features were extracted for the correlation analysis of overall survival (OS) and disease progression-free survival (DFS). We investigated cross-combination of 10 re-sampling methods and 4 classifiers for radiomics-based survival prediction.

Re-sampling techniques including oversampling and hybrid sampling achieved significant improvement on the area under the receiver operating characteristic curve (AUC) and prediction of the minority class in terms of precisions, recalls, and F-measures in both cohorts, as compared to no oversampling (Wilcoxon signed rank-sum test, p < 0.05). We observed that the combination method ADASYN oversampling+XGboost classifier (AUC of 0.70, accuracy of 0.74, sensitivity of 0.75, and specificity of 0.70) presented the highest performance for DFS prediction. ADASYN+SVM performed best (AUC of 0.80, accuracy of 0.76, sensitivity of 0.74, and specificity of 0.86) for OS prediction in our NPC cohort.

We identified optimal machine learning methods for the prediction of prognostication in NPC, which enhanced the applications of radiomics in precision oncology and clinical practice. Of note, re-sampling techniques showed a significant positive impact on prediction performance in imbalanced datasets.

Our study is limited by retrospective datasets in terms of relatively small numbers of instances.

n/a

No funding was received for this work.

Purpose:

We compared the parameters of gamma distribution (GD), intravoxel incoherent motion (IVIM), and mono-exponential (ME) models based on the goodness-of-fit, correlations, and effectiveness in the differential diagnosis among various orofacial lesions.

57 patients underwent 3-Tesla MRI examinations including TSE-DWI with 6 b-values (0-1000 s/mm²). The shape parameter k, scale parameter θ, and fractions of diffusion (ƒ₁, ƒ₂, and ƒ₃, which represent the cell component, parenchyma, and perfusion component, respectively) were obtained by GD model. D, D*, and f were obtained by IVIM model. ADC was obtained by ME model. Akaike information criterion was used to evaluate the goodness-of-fit. Correlations of the three methods were then assessed. All parameters were compared among inflammatory lesions (n=5) and benign (n=14) and malignant tumours (n=38).

There was no significant difference between the GD and IVIM models, but both models showed a better goodness-of-fit than the ME model (p<0.0001). The ƒ₁ showed strong negative correlations with D (ρ=–0.906) while the ƒ₃ showed moderately positive correlations with f (ρ=0.583). The malignant tumours showed significantly lower k, D, and ADC values (p<0.0001), which reflected their higher cellularity (higher ƒ₁ values) of benign tumours. No significant differences were found in any parameters between inflammatory lesions and either type of tumour.

The GD and IVIM models displayed comparable correlations and a better goodness-of-fit than the ME model. The GD model clearly showed that malignant tumours had higher cellularity than did benign tumours, which was attributed to their lower D and ADC values.

There was a limited number of benign tumours included.

This study was approved by the institutional ethics committee.

This research is supported by a JSPS KAKENHI (C) 18K09770.

Purpose:

To determine the diagnostic accuracy (AC) of multiparametric MRI (mp-MRI), fine-needle aspiration cytology (FNAC) and their combined use in differential diagnosis of parotid tumours.

We evaluated 68 benign (37 pleomorphic adenomas [PA], 23 Warthin tumours, 2 cysts [Cy], and 6 other benign lesions) and 21 malignant lesions (16 carcinomas, 5 lymphomas [L]). 2 ADC thresholds were identified to distinguish PA/Cy (ADC>1.3) and L (ADC<0.6) from other lesions. Adenopathies (p=.003), infiltration of adjacent structures (p<.001), hypointense/heterogeneous T2 signal (p=.002), and type C TIC (p<.001) were significantly related to malignancy and used in the definition of diagnostic algorithm. Applied to the validation group, it showed SE=69.23%, SP=95.24%, PPV=81.81%, NPV=90.90%, and AC=89.09%. Adding anamnestic data, it showed SE=78.57%, SP=95.24%, PPV=84.61%, NPV=93.02, and AC=89.47%.

FNAC demonstrated SE=77.78%, SP=100%, PPV=100%, NPV=96%, and AC=96.49%, with non-diagnostic/indeterminate results in 28.75% of cases.

When FNAC was used to characterise indeterminate lesions after mp-MRI, we obtained SE=80%, SP=95.34%, PPV=85.71%, NPV=93.18%, and AC=91.38%.

Preoperative mp-MRI, associated with FNAC in doubt lesions, reaches a high diagnostic accuracy and should be proposed as a preoperative diagnostic work-up of parotid lesions.

A relatively small sample size.

Approved by the Institutional Review Board; each patient signed a specific written informed consent.

No funding was received for this work.

Purpose:

In this prospective study, the value of pre-treatment dynamic contrast-enhanced (DCE) and diffusion-weighted (DW) MRI-derived parameters, as well as their changes early during treatment, was evaluated for predicting disease-free survival (DFS) and overall survival (OS) in patients with loco-regionally advanced head and neck squamous cell carcinoma (HNSCC) treated with concomitant chemo-radiotherapy (cCRT) with cisplatin.

MRI examinations were performed in 20 consecutive patients with loco-regionally advanced HNSCC at baseline and after 10 Grays (Gy) of cCRT. Tumour apparent diffusion coefficient (ADC) and DCE parameters (volume transfer constant [K_trans], extracellular extravascular volume fraction [v_e], and plasma volume fraction [V_p]) were measured. Relative changes in parameters from baseline to 10 Gy were calculated. Univariate and multivariate Cox regression analysis were conducted. Receiver operating characteristic (ROC) curve analysis was employed to identify parameters with the best diagnostic performance.

None of the parameters was identified to predict for DFS. On univariate analysis of OS, lower pre-treatment ADC (p=0.012), higher pre-treatment K_trans (p=0.026), and higher reduction in K_trans (p=0.014) from baseline to 10 Gy were identified as significant predictors. Multivariate analysis identified only higher pre-treatment K_trans (p=0.026; 95% confidence interval (CI): 0.000–0.132) as an independent predictor of OS. At ROC curve analysis, pre-treatment K_trans yielded an excellent diagnostic accuracy (area under curve [AUC]=0.95, sensitivity 93.3%, and specificity 80%).

In our group of HNSCC patients treated with cisplatin-based cCRT, pre-treatment K_trans was found to be a good predictor of OS.

A small sample size and high number of parameters to analyse.

At Ministry of Health, Republic of Slovenia No.22k/03/13.

No funding was received for this work.

Purpose:

To determine the utility of compressed sensing (CS) for head and neck MR imaging obtained with parallel imaging (PI) in patients with different diseases.

30 consecutive patients with various head and neck diseases underwent T2-weighted imaging by a 3T MR system (Vantage Galan 3T, Canon Medical Systems, Otawara, Japan) by PI with and without CS. On the quantitative assessment of the image quality, signal-to-noise ratio (SNR), percentage of the coefficient of variation (%CV), and contrast-to-noise ratio (CNR) in each patient were determined by ROI measurement. For qualitative assessment, two board-certified radiologists visually evaluated overall image quality, artefacts, and diagnostic performance by a 5-point scoring system. To compare quantitatively assessed image qualities, all indexes were compared between PI with and without CS by t-test. Interobserver agreements on each method were assessed by kappa statistics. To compare qualitative scores, a Wilcoxon’s singed-rank test was performed. Finally, mean examination time was compared between PI with and without CS.

SNR, %CV, and CNR of PI with CS were significantly better than those of PI without CS (p<0.0001). Interobserver agreements were assessed as significant and substantial (overall image quality: 0.67<κ<0.71, artefact: 0.65<κ<0.81, diagnostic performance: 0.62<κ<0.73). All qualitative image qualities had no significant differences (p>0.05). Examination time of PI with CS (83s ± 11s) was significantly shorter than without CS (173s ± 54s, p<0.0001).

Compressed sensing combined with parallel imaging is more useful than parallel imaging for head and neck MR examination in patients with different diseases.

A limited study population.

This study was a retrospective study and written informed consent was waived. This study was approved by the institutional review board of Fujita Health University Hospital.

This study was technically supported by Canon Medical Systems Corporation.

Purpose:

To evaluate the diagnostic accuracy of low-tube voltage 80-kV (peak) neck CT acquisition in the loco-regional extension of head and neck cancer, analysing the impact of ASIR-V algorithm on signal, noise, and image quality.

Two radiologists (with 5 and 10 years experience in head and neck cancer) separately analysed standard 120-kVp neck CTs and anonymised 80-kVp neck CTs of 20 patients with head and neck cancer who underwent CT (Revolution CT; GE Healthcare, Milwaukee, WI). 10 patients (6 oral cancers, 4 laryngeal cancers) who underwent surgery were unrolled. Images were reconstructed using an adaptive statistical iterative reconstruction (ASIR)-V algorithm on image quality. Image noise, the signal-to-noise ratio (SNR), and the contrast-to-noise ratio (CNR) were calculated for the lesions and were compared between the different post-processing ASIR-V % used. Reviewers were blinded to histology results. Sensitivity (SE), specificity (SP), positive predictive value (PPV), and negative predictive value (NPV) were calculated for standard 120-kVp CT and 80-kVp CT evaluations.

Low-tube voltage 80-kVp neck CT acquisition (SE 94%, SP 92.2%, VVP 95.1%, and NPV 90.3 %) is more accurate than 120-kVp CT (SE 85.7%, SP 77.8%, VVP 85.7 %, and NPV 77.8%) in the definition of the loco-regional extension in head and neck cancer. Compared to ASIR-V 0%, image reconstructed with ASIR-V demonstrated a significant reduction of image noise in 80-kVp CT images (P <0.05).

Low-tube voltage 80-kVp neck CT acquisition and reconstruction with ASIR-V could provide a better definition of the loco-regional extension in head and neck cancer.

The small cohort and the monocentric experience. Results must be validated in a large cohort.

The study was approved by an ethics committee.

No funding was received for this work.

Purpose:

To evaluate the prognostic value of radiomic features extracted from pre-treatment PET-CT images of the primary tumours (PT) and lymph nodes (LNs) in locally advanced nasopharyngeal carcinoma (NPC) patients.

A total of 145 consecutive patients (median age 48 years, 75.2% male) with newly diagnosed NPC were included. Radiomics analysis was performed on 118 PT and 66 LN images. Conventional PET parameters (SUV_max, SUV_mean, and TLG), shape-based features, and 6 robust texture features were extracted. This data was used to analyse the correlation with 3-year overall survival (OS) and progression-free survival (PRFS) based on a Kaplan-Meier log-rank test.

One PET radiomics feature (LN sphericity) was significantly predictive of OS (p=0.024) and PRFS (p=0.002). Independent CT predictive radiomic features were PT GLRM LRE (p=0.011), LN sphericity (p=0.011), and compactness (p=0.042) for OS, and PT GLRM LRE (p = 0.039) and LN sphericity (p = 0.008) for PRFS.

CT shape-based features and texture features extracted from primary tumours and LNs provide significant predictive value for survival in patients with NPC. LN PET radiomics information was significantly predictive for survival and performed better than primary features. The unity of LN radiomics can be further validated by prospective randomised trials, considering their potential in predicting clinical outcomes of tumour patients.

A retrospective study and small sample size.

This study has been approved by IRB.

No funding was received for this work.